Time-Stamped Annotations in High-Frequency Physiological Data: Evaluating an Approach for Assessing Annotation...

High-resolution neuromonitoring data with time-stamped clinical annotations offer valuable insights into treatment responses, though reliability is limited by manual documentation. This study aims to develop and demonstrate a three-step methodological framework to evaluate the plausibility of time-stamped clinical annotations using a high-frequency dataset. The framework integrates visual inspection of reference annotations and automated classification based on predefined physiological criteria.

Annotated interventions in the High-Resolution Collaborative European Neuro Trauma Effectiveness Research in Traumatic Brain Injury (CENTER-TBI) dataset were retrospectively analysed. Step 1 included visually inspecting physiotherapy and suctioning annotations to assess the overall plausibility of annotations at the patient level. In Step 2, an intracranial pressure (ICP)-based classification was applied to the period before osmotherapy annotations. Rejected annotations were those without sustained intracranial hypertension (ICP > 20 mm Hg for ≥ 5 min) beforehand. Step 3 classified the accepted annotations as effective (ICP reduction ≥ 10 mm Hg or normalisation) or ineffective on the basis of the post-annotation ICP trend.

Across 205 patients, 15,455 annotated interventions were identified. Visual inspection (Step 1) classified 90.2% of files as having moderate or high evidence of an annotation-signal relationship and 9.8% as having low evidence. The automated analysis (Step 2) identified 388 osmotherapy annotations in 76 patients; 140 (36.1%) were rejected, and 248 (63.9%) were retained. Step 3 found that, among valid events, 67.7% were effective and 32.3% ineffective. Rejected events were more frequent in low-evidence files (57.3%%) than in high-evidence ones (18.0%, p < 0.001).

This study provides the first systematic evaluation of time-stamped annotations in the CENTER-TBI high-resolution dataset. Concordant findings obtained using visual inspection and automated classification supports the credibility of both approaches and illustrates a framework for evaluating physiological plausibility, demonstrated here for osmotherapy but applicable to other annotated interventions. External validation is needed to prove the generalisability of the proposed algorithm.